OpenAI on AWS Bedrock: what it costs and what actually changed

A seven-year exclusivity deal ended in two days

Since 2019, every API call to OpenAI ran through Microsoft Azure. That was the arrangement: Microsoft invested roughly $13 billion total and got exclusive distribution rights for OpenAI's frontier models. Other clouds could offer OpenAI's open-weight releases, but not GPT-5.4 or anything built on proprietary training.

On April 27, 2026, Microsoft and OpenAI renegotiated. The exclusivity dropped. Microsoft keeps a nonexclusive license through 2032 and roughly 27% equity. OpenAI can now sell products on any cloud. The deal also reshuffled the economics: Microsoft stops paying OpenAI a revenue share; OpenAI continues paying Microsoft (capped) through 2030.

AWS had been positioning for this since February, when Amazon announced a $50 billion investment in OpenAI. (That is a different investment from Amazon's separate multi-billion dollar in Anthropic. OpenAI and Anthropic are competing companies.) Within 24 hours of the exclusivity announcement, AWS held an event in San Francisco and launched three OpenAI products on Bedrock: frontier models in limited preview, Codex on Bedrock, and Bedrock Managed Agents.

Worth noting the timing: Sam Altman attended the AWS announcement via video only. He was in Oakland at the Musk v. Altman trial the same week.

The honest answer about price

If you came here looking for a side-by-side "Bedrock is X% cheaper" number, there isn't one yet for frontier models. GPT-5.4 and GPT-5.5 on Bedrock are limited preview. AWS has not published pricing for them.

The reasonable expectation is price parity. When AWS added Anthropic Claude to Bedrock, they used pass-through pricing - same per-token rate as the direct Anthropic API. That is the most likely outcome here too. OpenAI controls the model economics; AWS adds infrastructure and takes its margin through compute, not a markup per token on top.

There is one place where Bedrock can be cheaper: AWS EDP commitments. If your team has negotiated an enterprise discount on AWS spend, Bedrock usage counts toward that commitment. A 15% EDP discount on $1,875/month of GPT-5.4 usage works out to $281 in savings. That adds up at scale. But it is not a Bedrock price cut - it is a discount on your total AWS bill that you were probably already getting on other services.

One thing that is definitely priced on Bedrock: the OSS models. gpt-oss-20b at $0.07/$0.30 and gpt-oss-120b at $0.15/$0.60 per million tokens are published and generally available. These have been on Bedrock since August 2025. The April 28 announcement was about the frontier models joining them.

What a GPT-5.4 workload costs either way

For a team running 10 million tokens a day on GPT-5.4 with a 7:3 input-to-output split: 7 million input tokens at $2.50/M is $17.50, 3 million output tokens at $15/M is $45. That puts the daily bill at $62.50 and the monthly at $1,875. Here is how the routing options compare:

10M tokens/day (7M input, 3M output) on GPT-5.4. Cache scenario uses $0.25/M cached input rate. Bedrock frontier pricing unconfirmed; EDP discount is illustrative for teams with existing AWS commitments.

The gpt-oss-120b row is there to show the range, not as a recommended swap. At 120B parameters it is not comparable to GPT-5.4 in quality, and saving $1,789/month is not free - you would need to run proper evals to know what breaks. For workloads that are overbuilt on frontier models, the number is still worth looking at.

What Bedrock actually adds

If per-token pricing is the same, the case for Bedrock is about infrastructure fit, not cost. The reasons are real but they matter more at organizational scale than for small teams.

Authentication is the most immediate practical win. On direct OpenAI, you manage API keys - rotating them, restricting them, distributing them across environments. On Bedrock, you use IAM. Your existing AWS role policies control which Lambda functions, ECS tasks, or EC2 instances can call which models. If you already have a mature IAM setup, adding an OpenAI model call is the same as adding any other Bedrock model call.

Data residency is the bigger deal for regulated industries. Inference on direct OpenAI runs on OpenAI's infrastructure. Inference on Bedrock runs within your chosen AWS region. For workloads that touch HIPAA-regulated patient data, GDPR-scoped personal data, or anything under FedRAMP, where the request actually executes has a different answer.

You can also layer Bedrock Guardrails on top. Content filtering, PII redaction, topic blocking - the same guardrail layer that works with Claude or Llama 4 on Bedrock applies to OpenAI models too. For teams that have already built guardrails workflows, this means OpenAI models slot into the existing setup without rebuilding the safety layer.

The last thing worth naming: unified billing. OpenAI costs show up on your AWS invoice. This sounds like a minor accounting detail, but in larger orgs where procurement and compliance have AWS approved as a vendor, getting a separate OpenAI contract through that process can take months. Bedrock routes around that. We've talked to teams that waited 3-plus months for InfoSec to approve a new API vendor, while their AWS account was already fully approved. That friction is real, and it is not unique to OpenAI.

Amazon Bedrock Managed Agents

The third product in the April 28 announcement was Bedrock Managed Agents powered by OpenAI. Also limited preview. Pricing not published.

What it is: a managed agent orchestration service that pairs OpenAI frontier models with AWS infrastructure. Each agent has its own identity in your AWS environment, logs its actions, and runs statefully - meaning it maintains context across sessions. This was the technology at the center of the Microsoft-OpenAI dispute. Microsoft initially argued that the stateful runtime was covered by their API exclusivity. The April 27 renegotiation resolved that.

The comparison to OpenAI's Assistants API is obvious and the technical differences are real. The Assistants API sends everything to OpenAI's servers. Bedrock Managed Agents keeps execution within your VPC. For the same reasons data residency matters for inference, it matters for agent state and tool call logs.

Pricing will likely include two components: token costs at the OpenAI rate, plus some fee for session management and tool calls. That follows the Assistants API pattern. Until AWS publishes numbers, any estimate is speculation.

When the direct API still makes more sense

GPT-5.4 Mini and GPT-5.4 Nano are not on Bedrock. Those two models at $0.75/$4.50 and $0.20/$1.25 per million tokens are direct-API-only as of today. For teams running classification, summarization, or lightweight tasks at scale, Bedrock is not an option without switching to a different model tier.

The frontier models are still limited preview. Getting access requires an application through AWS; general availability timing has not been announced. If you need GPT-5.4 for production right now, the direct API is the path without a waitlist.

And for small teams or solo developers: if you have no AWS EDP commitments and no compliance requirements, Bedrock adds complexity without financial benefit. An API key and a credit card is a simpler setup than IAM roles and Bedrock endpoints. The infrastructure benefits only matter at organizational scale. The answer to "should I switch to Bedrock for OpenAI models" is mostly "it depends on whether you are already deep in AWS."

Sources

• - AWS What's Next event announcement (April 28, 2026): aws.amazon.com
• - TechCrunch: Microsoft exclusivity ends (April 27, 2026): techcrunch.com
• - TechCrunch: OpenAI products on AWS (April 28, 2026): techcrunch.com
• - The Decoder: OpenAI lands on AWS details: the-decoder.com
• - AWS Bedrock pricing (official): aws.amazon.com/bedrock/pricing
• - AWS Bedrock supported models: docs.aws.amazon.com
• - TechCrunch: First OpenAI OSS models on AWS (August 2025): techcrunch.com
• - TechCrunch: Amazon $50B OpenAI deal (February 2026): techcrunch.com